This article by Nina Xiang originally appeared on China Money Network, the best data intelligence platform tracking China’s tech and venture capital markets (access requires subscription).

Entering 2019, the buzzword in the artificial intelligence industry has taken a decidedly new tone. “Bottleneck” has been much talked about, from opinion articles questioning if deep learning “will be the best long-term solution” for building intelligent machines,” to startup founders openly discussing the limitations of its future potentials.

The latest such reflection came from Yuan Peijiang, founder of Chinese AI startup SensingTech. “One of the biggest challenges for artificial intelligence is that we have reached a bottleneck of the current technology, and it’s difficult to achieve more breakthroughs from it,” Yuan, a Tsinghua graduate and a Ph. D [holder] in electronics and computer engineering from the University of Western Ontario, told China Money Network at SensingTech’s Beijing office last month. “From the capital markets’ perspective, there will be a bottleneck in terms of returns. That means capital investments will decrease and the industry’s growth will slow.”

Such sobering remarks echo the broader macroeconomic uncertainties and clear signs that tech companies are bracing for a much tougher operating environment this year. Since the end of 2018, Chinese tech firms including Didi, Baidu,, Huawei, Sina, Mobike, iFlytek, Qudian and others have taken steps to lay off staff, or “implement operational optimizations,” as some companies termed it.

For AI startups, the type of breakneck growth experienced in the past couple of years is not expected to continue this year. SenseTime, among the highest valued AI companies in the world with a reportedly price tag of $6 billion, did not set objectives to “double staff” and “open more offices” in its annual year-end company gathering this year. The company touted such achievements in the past two years as a way to build up morale, according to someone with direct information of the matter. This sense of humility is echoed at other AI companies as well. Megvii, also known as Face++, is locating new staff in less fancy offices, instead of its main base in Raycom Infotech Park, an expensive luxury office space in central Zhongguancun in Beijing.

For the venture capital market, runaway massive fundraising deals and sky-bound valuation growth are expected to moderate. Though, big deals will still take place, but will concentrate on established players with entry barriers and core tech competency, Yuan reckons. A review of mega VC deals over $100 million in the Chinese AI space reveals that 2018 stood out as a record-setting year that is unlikely to be matched this year or in the near future. The window for setting up leading AI companies my be closnig, and it will be increasingly difficult for new AI startups to raise capital, Yuan said.

(Image credit: China Money Network)

In 2018, there were 22 such mega deals, compared to seven in 2017 and four in 2016. In 2019, there were only one such deals so far, which is AI chip maker Horizon Robotics raising a US$600 million series B round in February. The fundraising pace has slowed to one mega deal in two months from the searing speed of almost two mega deals each month last year.

Not so cold after all

Yuan was a deputy professor at Beihang University since 2009, participating in the development of high performance bionic quadruped robots for Chinese military and aerospace manufacturing robots. Before that, he was an assistant researcher at Tsinghua University studying topics of facial recognition and smart robots.

In 2016, he co-founded SensingTech with a former colleague at Beihang Univeristy, Shi Zhenyun, to seek opportunities in facial recognition’s commercial application. Compared to the “Big Four” in China’s facial recognition field, SenseTime, Megvii, Yitu Technology and CloudWalk, which were founded in 2014, 2011, 2012 and 2015, SensingTech was a latecomer.

But the company was able to secure a piece of the public security market, obtaining government contracts in provinces including Gansu, Guiyang, Shanxi, Henan and Yunnan. The company said that its systems have also been used in important forums such as the 19th National Congress and the Forum on China-Africa Cooperation to ensure on-site security.

Compared to other facial recognition companies like the “Big Four” that have raised billions of dollars, SensingTech raised a modest amount of capital. In 2018, it raised a Series B round of financing worth several hundred million yuan (RMB 100 million is worth around $15 million), as the company refused to disclose specific numbers. A year earlier, it secured around RMB 100 million ($15 million) Series A round.

Though the size of the capital was not large, the investors were largely bluechip Chinese state-backed funds include CASH Capital, which counts Chinese Academy of Sciences Holdings Co., Ltd. as its anchor limited partner. SDIC Capital Co., Ltd., which is affiliated with the State Development and Investment Corporation (SDIC), is also an investor of SensingTech. This gives the company unique access and informational edge on government policy directions.

Faced with the challenging macroeconomic environment, Yuan said the company had already adjusted strategies to deal with the harsh conditions. “We are in a ‘winter’, and the best strategy is to strengthen yourself at this time,” Yuan said. “[We are growing] our management capabilities, our execution skills, our nimbleness to meet market demand, and our team’s operational efficiency.”

Though everyone will feel more pressure to deliver this year, Yuan is confident that “this winter” won’t be so cold after all. “This industry will run into real trouble if there is a lack of confidence and support. But we continue to see confidence from the government to support AI,” Yuan said, referring to recent positive interactions with all levels of governments in China. “We can clearly feel that many local governments still very much welcome companies like us for cooperation. This confidence is the most precious [for the sector].”

Looking at the industry’s growth in the long term, Yuan predicts that there will be the emergence of around five AI behemoths like today’s Alibaba and Tencent, or perhaps worth even more than them, in around 20 years. Because AI is a winner-take-all sector and dominating companies will absorb most of the resources and market share, this creation of next generation AI giants will evolve similarly to how current internet giants were formed.

Technology bottlenecks

In a recent article, Alan Yuille, a professor of cognitive science and computer science at Johns Hopkins University, summarized the bottlenecks of deep learning in its application in enabling machines to “see” the world.

These bottlenecks include the way deep learning is designed to work for specific tasks, its dependence on large annotated data sets for training and testing, and its poor performance in real-world scenarios. For example, it is very easy to trick deep learning algorithms.

On the other hand, for areas where deep learning has been very successful in, the most significant benefits have been reaped. For instance, facial recognition software got 20 times better at searching a database to find a matching photograph between 2014 to 2018, according to the National Institute of Standards and Technology’s (NIST) in the US. That pace of improvement is impossible to repeat in the future as the ratio of success rate in finding a match is already closer to 99.8%, says NIST.

“The more we are toward the end, it will cost a lot more to achieve even the tiniest improvements. It is like climbing a mountain, the higher you are, the more difficult it is,” Yuan said. But he believes there are still potential in applying AI in different industries to improve efficiencies and save costs. “It is similar to the invention of internal combustion engines. There are still lots of opportunities to create value in applying it in all sorts of scenarios.”

For SensingTech, the areas it is exploring new applications in include education and residential security. In education, facial recognition and other technologies can be used for student safety, crowd management, classes coordination, for example. Similarly in ensuring residential community’s safety.

But it won’t be easy. A recent leaked procurement document by a school in Guangdong province has created an uproar. The school wants to purchase smart wristbands for students to monitor their movements, activities, in-class behavior and even their shopping history, prompting many to voice concern that it is a dangerous invasion of privacy in students’ lives. Such public opposition is just one of the challenges SensingTech has to overcome to survive the “winter” and thrive.

Leave a comment

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.